Moral values are fundamental principles and convictions that guide how people act and relate to others, influencing their daily ethical decisions and behaviours. Understanding the limitations in the detection of these values is crucial, especially in digital media, where the interpretation of morality can be more complex due to the diversity of content and context. Digital platforms have transformed the way people communicate and interact, creating a greater need to ensure that the content shared is appropriate. Identifying these values can allow for a better understanding of the intentions and underlying messages in content, making people more aware of how it can influence perception and decision-making. This work focuses on the evaluation of the performance of models based on BERT and RoBERTa transformers, which represent the state of the art in natural language processing (NLP) in a variety of applications.In particular, the ability of these models to detect morality in texts is investigated using the ethical foundations defined in Moral Foundations Theory (MFT), which identifies five moral traits and distinguishes between vice and virtue. It also analyses different levels of complexity in morality detection and explores the impact on the models of incorporating subjective information and additional detail through the use of lexical resources reflecting emotion and morality. Finally, it examines how these approaches perform in different domains and how they benefit text comprehension. Results show that the addition of these lexicons, despite depending on the complexity of the task, positively influences the models’ ability to distinguish the underlying morality in the text. Results improvement is observed in both situations, similar to the training do- main and in different domains, demonstrating the effectiveness of approaches that integrate enriched data with subjective perspectives to increase the models’ robustness.